Related Experiment Video
Updated: Apr 26, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.6K
FedSemiDG: Domain generalized federated semi-supervised medical image segmentation
Zhipeng Deng1, Zhe Xu2, Tsuyoshi Isshiki3
1Medical Artificial Intelligence Lab, Westlake University, Hangzhou, China; Department of Information and Communication Engineering, School of Engineering, Institute of Science Tokyo, Tokyo, Japan.
Medical Image Analysis
|April 24, 2026
Summary
Federated semi-supervised learning (FSSL) struggles with domain shifts in medical imaging. Our novel framework, FGASL, improves generalization to unseen domains by using adaptive aggregation and refined pseudo-labels for better multi-center collaboration.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Machine Learning
Background:
- Federated semi-supervised learning (FSSL) trains models on decentralized data without sharing raw information, addressing privacy concerns in medical imaging.
- Domain shift, a significant challenge in FSSL, leads to suboptimal performance when models encounter data from different sources or scanners.
- Existing FSSL methods often fail to adequately address the domain generalization problem, limiting their effectiveness in real-world, multi-center medical applications.
Purpose of the Study:
- To introduce and evaluate a novel framework for domain generalized federated semi-supervised learning (FedSemiDG) in medical image segmentation.
- To address the under-explored challenge of domain shift within the FSSL paradigm for medical image analysis.
- To develop a method that enables models to generalize robustly to unseen medical imaging domains using limited labeled and abundant unlabeled data.
Main Methods:
- Federated Generalization-Aware Semi-Supervised Learning (FGASL) framework integrating global and local strategies.
- Global level: Generalization-Aware Aggregation (GAA) adaptively weights local models based on generalization performance.
- Local level: Dual-Teacher Adaptive Pseudo Label Refinement (DR) and Perturbation-Invariant Alignment (PIA) enhance pseudo-label reliability and feature consistency.
Main Results:
- FGASL significantly outperforms state-of-the-art FSSL and domain generalization approaches across four medical segmentation tasks.
- The proposed method demonstrates robust generalization capabilities on unseen medical imaging domains.
- Experiments on cardiac MRI, spine MRI, bladder cancer MRI, and colorectal polyp segmentation validate the framework's effectiveness.
Conclusions:
- FGASL provides a practical and effective solution for domain shifts in federated semi-supervised learning for medical image segmentation.
- The framework advances multi-center collaboration by enabling robust model generalization in privacy-sensitive healthcare applications.
- This work addresses a critical limitation in current FSSL methods, paving the way for more reliable AI in diverse clinical settings.
